Pyramid-JiT: New Decoder-Only Diffusion Architecture Cuts Image Training Cost 4.3× vs. Prior Model

Researchers have proposed Pyramid-JiT (P-JiT), a decoder-only pixel-space diffusion architecture that generates images at multiple resolutions along the DiT trunk, eliminating the need for a separate VAE. The model reaches the FD-DINOv2 quality score of the team's prior Linum v2 latent diffusion model on 11.3× fewer training samples and in 4.3× fewer GPU-hours at 4× the pixel count — a meaningful efficiency leap on the path toward affordable video generation.

The work is positioned as a research stepping stone toward a future Linum v3 video model, with the explicit goal of slashing attention costs driven by large token sequences in video (Linum v2 required 110K tokens for a short 720p clip). Code and weights are released under the Apache 2.0 license. The team frames P-JiT as a testbed architecture rather than a full product release, inviting the broader research community to explore similar token-reduction strategies for efficient video generation.